Deep transfer learning model in predicting the longitudinal wind pressure time series on a high-rise building

Abstract

Accurately obtaining the spatial and temporal characteristics of wind loading is key to the wind resistance of slender structures. This paper proposes a deep-transfer-learning-based framework TL-POD-LSTM that combines transfer learning (TL) with proper orthogonal decomposition (POD) and long short-term memory network (LSTM). TL-POD-LSTM predicts the pressure time series at any longitudinal location using data from very few sensors by training the source domain model, transferring the source model to the untrained target model, training and fine-tuning the target model, and predicting the loading using the target model. The wind tunnel experimental data of pressure time series at four laps of pressure taps in various longitudinal locations on a square cylinder was used to compare TL-POD-LSTM with the previous POD-LSTM model that uses only the target domain data. Various combinations of longitudinal spacings, training tap schemes, and wind incidences were tested. The total error, local error, pressure statistics, aerodynamics, and error factors were analyzed. TL-POD-LSTM significantly outperforms POD-LSTM in precision and generalization performances. Using only 4 taps at the target domain, TL-POD-LSTM improves the determination coefficient from 0.194 to 0.976 at an incidence of 45° and a longitudinal spacing equal to the side length. The precision of TL-POD-LSTM has no relevance to the longitudinal spacing between the source and target domains. The source domain data should be carefully selected to reduce both the source model errors due to adjacent-tap data differences and the transfer learning errors due to source-target mode vector differences.